-
Notifications
You must be signed in to change notification settings - Fork 32
/
Copy path18-explaining_models_and_predictions.Rmd
222 lines (162 loc) · 8.05 KB
/
18-explaining_models_and_predictions.Rmd
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
# Explaining models and predictions
**Learning objectives:**
- Recognize some R packages for model explanations.
- Use {DALEX} and {DALEXtra} to produce local model explanations for a model trained using {tidymodels}.
- Use {DALEX} and {DALEXtra} to produce global model explanations for a model trained using {tidymodels}.
- Use {DALEX} and {DALEXtra} to produce partial dependence profiles for a model trained using {tidymodels}.
## Chapter 18 Setup
Load in the data and set up explainer
```{r 18-load-data, warning=FALSE, message=FALSE}
library(tidymodels)
library(skimr)
library(DALEX)
library(DALEXtra)
library(iBreakDown)
rush_model <- readRDS(here::here("data", "18-fit_rush_yards.RDS"))
rush_df <- readRDS(here::here("data", "18-nfl_rush_df.RDS"))
skim(rush_df)
explainer_boost <-
explain_tidymodels(
rush_model,
data = rush_df,
y = rush_df$rushing_yards,
verbose = TRUE
)
```
## Overview
## Local Explanations
- Provides information about a prediction for a single observation
- Which variables contribute to this result the most?
- "Break-down" explanations compute the contribution from each feature
- Results for many explanatory variables can be presented in a limited space
- Only the additive attributions, misleading for models with interactions
- Break-down plots with interactions
- More accurate if the model itself uses interactions
- Much more time-consuming
- Interactions is not based on any formal statistical-significance test
- SHapley Additive exPlanations (SHAP) are based on “Shapley values”
- "Cooperation is beneficial, because it may bring more benefit than individual actions"
- Decompose a model’s predictions into contributions that can be attributed additively to different explanatory variables
- If the model is not additive, then the Shapley values may be misleading
```{r 18-local, eval=FALSE}
#Break-down
boost_breakdown <- predict_parts(explainer = explainer_boost,
new_observation = sample_n(rush_df,1))
png(file="images/18_boost_breakdown.png", width = 600)
plot(boost_breakdown)
dev.off()
#Break-dwon Interactions
boost_breakdown2 <- predict_parts(explainer = explainer_boost,
new_observation = sample_n(rush_df,1),
type = "break_down_interactions")
png(file="images/18_boost_breakdown2.png", width = 600)
plot(boost_breakdown2)
dev.off()
#SHAP
boost_breakdown3 <- predict_parts(explainer = explainer_boost,
new_observation = sample_n(rush_df,1),
type = "shap")
png(file="images/18_boost_breakdown3.png", width = 600)
plot(boost_breakdown3)
dev.off()
```
## Local Explanations for Interactions
- "Ceteris-paribus" profiles show how a model’s prediction would change if the value of a single exploratory variable changed
- Graphical representation is easy to understand and explain
- Not a valid assumption with highly correlated or interaction variables
```{r 18-ceterus, eval=FALSE}
#Ceterus Paribus
boost_paribus <- predict_profile(explainer = explainer_boost,
new_observation = sample_n(rush_df,1),
variables = c("rusher_age", "yardline_100"))
png(file="images/18_boost_paribus.png")
plot(boost_paribus, variables = c("rusher_age"))
dev.off()
png(file="images/18_boost_paribus2.png")
plot(boost_paribus, variables = c("yardline_100"))
dev.off()
```
## Global Explanations
- Which features are most important in driving the predictions aggregated over the whole training set
- Measure how much does a model’s performance change if the effect of a selected explanatory variable(s) is(are) removed
- If variables are correlated, then models like random forest are expected to spread importance across many variables
- Dependent on the random nature of the permutations
```{r 18-global, eval=FALSE}
boost_vip <- model_parts(explainer_boost, loss_function = loss_root_mean_square)
png(file="images/18_boost_vip.png")
plot(boost_vip, max_featuers = 10)
dev.off()
```
## Global Explanations from Local Explanations
- Partial-dependence plots
- How does the expected value of model prediction behave as a function of a selected explanatory variable?
- PD profiles are averages of CP profiles
- Problematic for correlated explanatory variables
```{r 18-global_profile, eval=FALSE}
boost_profile <- model_profile(explainer_boost,
N = 1000,
variables = "rusher_age",
groups = "position")
png(file="images/18_boost_profile.png")
plot(boost_profile)
dev.off()
```
## References
[DALEX Github](https://modeloriented.github.io/DALEX/)
[DALEXtra Github](https://github.com/ModelOriented/DALEXtra)
[Exploratory Model Anaylsis](https://ema.drwhy.ai/)
## Meeting Videos
### Cohort 1
`r knitr::include_url("https://www.youtube.com/embed/jk6hHNyrcSo")`
<details>
<summary> Meeting chat log </summary>
```
00:03:12 tan_iphone: Hullo!
00:05:15 tan_iphone: I seem to have influenced the ballot ok tho!
00:06:20 tan_iphone: Boaty mc boat dog
00:13:56 Jon Harmon (jonthegeek): https://cran.r-project.org/package=nflfastR
00:17:17 Tony ElHabr: Acronyms are my fav
00:17:21 Tony ElHabr: *bacronyms
00:17:31 Jon Harmon (jonthegeek): I think it stands for "I'm a Doctor Who fan."
00:17:59 Jordan Krogmann: they have the death bots as the logo
00:18:10 Jordan Krogmann: or whatever... I am a casual doctor fan
00:18:28 Jon Harmon (jonthegeek): https://dalex.drwhy.ai/
00:24:26 Jon Harmon (jonthegeek): Rowers in a boat is the explanation I've heard a lot. You can't just be the fastest, you need to also be in sync with the other rowers.
00:24:52 Tony ElHabr: Is the default shapley? Or is it permutation? And what’s the difference?
00:31:19 Jon Harmon (jonthegeek): Dunno about anybody else but I had to look up what that meant: ceteris paribus = "all other things being equal."
00:31:38 Tony ElHabr: This is the same as partial dependence plots?
00:32:45 Tony ElHabr: Oh, CP is the instance-level equivalent of PDP (used for data-set level)
00:33:46 Jon Harmon (jonthegeek): "A profile showing how an individual observation’s prediction changes as a function of a given feature is called an ICE (individual conditional expectation) profile or a CP (ceteris paribus) profile."
Yeah, they never definitively explain PDP vs these other terms. I hate to make them rebuild this monster chapter but I think I have thoughts about making some stuff in here clearer!
00:37:52 Tony ElHabr: i don’t think vip supports catboost 😢
00:39:04 Jon Harmon (jonthegeek): Wait, do you pronounce it "vip" or "V I P"?
00:39:20 Tony ElHabr: Yeah v i p
00:39:46 tan_iphone: very important predictor
00:40:04 tan_iphone: But wait is there a prequel?
00:40:56 Tony ElHabr: meme (tweet) for the day: https://twitter.com/benbbaldwin/status/1431300791516663816?s=20
00:42:08 Jon Harmon (jonthegeek): Dude, you can't send something that requires audio in the middle of a meeting 🙃
01:05:41 tan_iphone: Lmaoooo I have not nearly read enough of the book to do it
01:05:47 tan_iphone: Currently at the gym :) cheers folks
```
</details>
### Cohort 3
`r knitr::include_url("https://www.youtube.com/embed/WXZ5DKaK5mQ")`
### Cohort 4
`r knitr::include_url("https://www.youtube.com/embed/XCtHncZwDB0")`
<details>
<summary> Meeting chat log </summary>
```
00:22:02 Federica Gazzelloni: this is interesting: “ game theory techniques like SHAP are able to provide insight into how much crime is “good,” and how much is “too much.” (https://geophy.com/insights/whats-the-shap-how-crime-influences-property-value/)
00:26:51 Federica Gazzelloni: example of using shap:
00:26:56 Federica Gazzelloni: https://juliasilge.com/blog/board-games/
```
</details>